A medical image classification method based on self-regularized adversarial learning.

Area of Science:

• Medical Imaging Analysis
• Deep Learning
• Artificial Intelligence in Healthcare

Background:

• Medical image classification faces challenges like small datasets and imbalanced classes.
• Generative adversarial networks (GANs) show promise for data augmentation in medical imaging.
• GAN performance often depends on high-quality generated images and large training datasets.

Purpose of the Study:

• To propose an adversarial learning-based classification framework (GAN-DL) for improved medical image classification.
• To utilize GAN models as supplementary regularization terms to address data limitations.
• To enhance classification performance in challenging medical imaging scenarios.

Main Methods:

• The GAN-DL framework includes a feature extraction network (F-Net), classifier, reconstruction network (R-Net), and discriminator network (D-Net).
• An iterative adversarial learning strategy with network-specific loss functions guides training.
• Loss functions act as regularization, derived automatically without additional data annotation.

Main Results:

• The GAN-DL framework demonstrated superior performance over 13 classic deep learning methods on COVID-19 and OPSCC datasets.
• Achieved high precision, sensitivity, specificity, and F1-score on both datasets.
• Ablation studies confirmed the critical role of the discriminator network (D-Net) and provided insights into methodology.

Conclusions:

• The adversarial-based framework enhances medical image classification accuracy and mitigates overfitting.
• The modular design offers flexibility for various clinical contexts and medical imaging applications.
• GAN-based regularization proves effective for improving deep learning in medical image analysis.